Building Virtual Cities:

3 downloads 0 Views 172KB Size Report
evaluation of a city generation system called The. Virtual City Builder (VCB). It then describes simulation experiments with VCB which uncover relationships.

Building Virtual Cities: applying urban planning principles to the design of virtual environments Rob Ingram and Steve Benford Department of Computer Science The University of Nottingham Nottingham NG7 2RD. UK. E-mail: { rji, sdb }@cs.nott.ac.uk

ABSTRACT

Urban planners have developed many theories of how urban design affects people’s experience of real-world cities. This paper considers how this body of work might be applied to the design of virtual environments. The paper presents and compares two pieces of research, each proposing algorithms for automatically enhancing or generating virtual environments and each inspired by a different theory of urban design. The first describes how Kevin Lynch’s work on the legibility of cities can be applied to improve the navigability of three dimensional information visualisations. The paper describes the design, construction and experimental evaluation of a system called LEADS which implements some general purpose legibility enhancing algorithms. The second describes how Hillier and Hansen’s work on the social logic of space has been applied to the construction of multi-user virtual cities. The paper describes the design, implementation and experimental evaluation of a city generation system called The Virtual City Builder (VCB). It then describes simulation experiments with VCB which uncover relationships between different city layouts and opportunities for navigation and social encounter. The paper concludes by comparing both pieces of work and by outlining possibilities for their future integration. KEYWORDS

Information visualisation, urban planning, cooperative work, virtual reality, legibility, navigation. INTRODUCTION

Urban planners have long since recognised that the design of urban environments affects many aspects of their use including their navigability and the degree to which they encourage social encounters between their inhabitants. This paper presents an initial exploration into how theories of urban design can contribute to the Permission to copy without fee all or part of this material is granted provided that the copies are not made or distributed for commercial advantage, the ACM copyright notice and the title of the publication and its date appear, and notice is given that copying is by permission of the Association for Computing Machinery. To copy otherwise, or to republish, requires a fee and/or specific permission.

John Bowers Department of Psychology The University of Manchester Oxford Road, Manchester UK. E-mail: [email protected]

design of virtual environments. We describe and compare two independent pieces of research which have been inspired by different theories of urban design. Both share a common approach in that they specify algorithms for the automatic design of virtual environments. The first considers how existing 3D information visualisations might be automatically enhanced to support people in learning to navigate them. The second considers how collaborative virtual environments might be automatically generated so as encourage ease of navigation and promote opportunities for social encounters. Both pieces of research include experimental work which provides initial evidence for their utility. In addition to describing specific algorithms for the design of a variety of virtual environments (both abstract and literal), the paper compares the approaches adopted, draws out directions for future work and, perhaps most importantly, indicates the potential of urban planning theory as a basis for research into the design of virtual environments. Each of the following descriptions (sections two and three below) begins with a statement of the specific goals to be addressed, then provides a quick overview of the urban theory which has provided the inspiration for the research, describes specific algorithms for the enhancement or construction of virtual environments and, finally, discusses initial experimental work. Following this, section four compares both pieces of work and draws out some more general conclusions. ENHANCING THE LEGIBILITY OF INFORMATION VISUALISATIONS

In 1960 Kevin Lynch published his seminal book The Image of the City [Lynch60]. This book developed a theory of urban design centred around the concept of legibility, a property of real-world cities which influences peoples’ ability to learn to navigate them. In fact, Lynch went further by proposing a set of design techniques to help urban planners to improve the legibility of cities. Navigation is also a key issue for information systems. Jacob Nielsen addressed the issue of navigation in hypertext systems in a series of papers, pointing to examples of users becoming lost (e.g. [Nielsen90]). This problem can only have increased in importance as a

result of the emergence of the World Wide Web as a global hypermedia data-base containing millions of pieces of information. In turn, authors such as Kim Fairchild have proposed that three dimensional visualisation techniques might increase the amount of information that people can meaningfully manage [Fairchild93]. However, our experience of navigation in current three dimensional visualisations, and in virtual environments in general, suggests that effective navigation remains a critical issue.

• Districts - sections of the environment which have a distinct character which provides coherence, allowing the whole to be viewed as a single entity. Districts may be identifiable by the nature of the architecture of their buildings or by their use.

Recognising the similarity between the problems of navigating real cities and navigating three dimensional representations of information, we therefore wondered whether Lynch’s principles could be adapted to virtual environments. Due to space constraints, the following sections present a only very brief summary of this work. Full details can be found in [Ingram96].

• Paths - major avenues of travel through the environment such as major roads.

• Landmarks - static and recognisable objects which can be used to give a sense of location and bearing. Examples of landmarks might be prominent buildings or monuments or, on a smaller scale, recognisable shopfronts or roadside installations.

• Nodes - important points of interest along paths, e.g. road junctions or town squares. • Edges - structures or features providing borders to districts or linear obstacles. Examples might be the waterfront in cities built on large rivers, or a major road. Note in the latter case that the road may have a dual nature, being a path for someone travelling in a car but an edge to the pedestrian.

Goal: helping people learn to navigate

At this stage, we must be careful to precisely state our goal. We are not primarily concerned with the mechanics of controlling viewpoints and embodiments; nor are we primarily concerned with how people navigate unfamiliar environments (i.e. ones that they haven’t seen before). Instead, our specific aim is to support people in better learning to navigate visualisations as a result of repeated exposures to them over a period of time. In other words, we are concerned with how people can be aided in gradually learning the structure of a graphical space. Consequently, we anticipate that the results of our work would be mostly applicable to visualisations which are persistent, which evolve relatively slowly in relation to their overall size and which are repeatedly visited. A three dimensional visualisation of a region of the World Wide Web would be a good example.

Lynch also specified a number of design techniques aimed at creating or enhancing these features within a given urban environment in order to improve its legibility. Implementation: legibility algorithms for information visualisations

We now consider how the general notion of legibility and the five specific legibility features identified by Lynch might be adapted for use in virtual environments. In particular, we specify some general purpose algorithms for automatically creating or enhancing legibility features within a range of pre-existing 3-D information visualisations. We also describe how these algorithms have been implemented in a prototype system called LEADS (LEgibility for Abstract Data Spaces). LEADS has been constructed as an example “legibility layer”, an independent sub-system which, following some configuration and basic integration, is able to post-process the output of other visualisation systems in order to enhance their legibility.

Theory: the legibility of urban environments

We now present a very brief review of Lynch’s concept of legibility. Lynch defines the legibility of a city as: “-the ease with which its parts may be recognised and can be organised into a coherent pattern- ”

Districts

Here, Lynch is referring to the formation of a cognitive map within a person's mind [Passini92], a structure which is an internal representation of an environment which they use as a reference when navigating to a destination. The Image of the City describes experiments carried out in a number of major US cities which suggest how cognitive maps are built up over time. The experiments involved obtaining information from long term inhabitants of the cities in the form of, for example, interviews, written descriptions of journeys through the city and drawn maps. By examining this data Lynch proposed that the following five elements of urban landscapes had been identified as the primary building blocks of cognitive maps:

The discovery of districts within the data is a matter of finding groups of items which have strong similarities to each other which they do not share with other items in the space. In order to identify districts within an arbitrary data space we can use techniques from the field of cluster analysis, for which a number of algorithms have already been developed. Once districts have been identified, they need to be represented in the visualisation. For example, one might use the attributes of colour and shape to give each district a distinct character (provided that these attributes had not already been assigned specific meanings within the visualisation). Districts also provide the basis for creating the remaining legibility features.

2

In choosing an algorithm for the initial LEADS implementation, we set the criteria that it should be simple, due to time constraints on the project, relatively computationally inexpensive and reasonably effective for a wide range of clusters. We have initially adopted Zahn’s Minimal Spanning Tree algorithm (MST) [Zahn]. The basic algorithm as described by Zahn is able to detect clusters of varying shapes and sizes so long as they are relatively distinct (some of the specific cases where it will not perform accurately are described in the original paper).

a better choice for the placement of landmarks. The third method places landmarks at the centre of the triangles formed by the centroids of any group of three adjacent clusters. This will produce approximately the same number of landmarks as the cluster intersection method. The position of the landmarks, although seemingly quite similar to that produced by the second is significantly different in that the landmark is placed in a central position between the points which most typify the districts rather than relying solely on the geometric point of intersection. We suspect that this will give a more even balance to the placement of the landmarks than simply positioning them at the intersection points and will make the landmark positions more responsive to the shapes of the clusters as well as their positions. This is the method that we have implemented in the LEADS system.

Edges

Edges are (usually) imposing features which section off one area from another. It seems sensible therefore to define edges as existing between adjoining districts. We have considered three possible methods for defining edges. The first is a quick method which might work tolerably well for those clusters which are generally spherical or cuboid in shape. This simple approach is to find the nearest neighbour data items between the clusters using the same similarity measure as was employed in the initial clustering process. The edge can then be placed between the clusters along the line connecting these two items, with an appropriate orientation. Provided that the objects do not excessively cut into the clusters or are positioned far from their logical joining point these methods should be effective in providing a reference point and defining the borders of clusters. The second method involves finding the hull of each district and creating an edge just beyond this. However, finding the hull of a cluster of objects in 3-D space may work out to be computationally expensive. The third method is a refinement of this which again involves finding the hulls of the adjoining districts. Once this is complete the edge could be defined by interpolation between the points along adjoining edges of districts. The initial implementation of the LEADS system utilises the first method described here, positioning the edges between cluster nearest neighbours.

Nodes and paths

We propose that paths should be composed of links between individual elements of the visualisation which would therefore act as nodes. Two main issues need to be considered. The first is which, if any, elements will be used as initial nodes and paths. The second is how the path layout might evolve with use. Assuming no prior usage information, it will be necessary to make educated guesses at data elements which might possibly become nodes through usage. We have implemented the following approach in LEADS: • The nearest neighbour elements between districts are initial choices as these are the most similar items across district boundaries. We call such nodes gateway nodes. • An additional main node for each district might be defined as that data element closest to the centroid of the cluster as this might be seen as being the item most typical of the district. In network type spaces there may be further data available to identify such nodes automatically, such as connectivity information (valency) or measures of amounts of data stored, both of which imply importance.

Landmarks

Landmarks need to be stable points within the data space in order to serve as a common reference for navigation. We have considered three possible methods for defining landmark positions. The first method places landmarks at the centroids of district clusters. Most partitional clustering algorithms will define the centroid of a group while it is being formed. This centroid is a virtual data element that best describes the cluster as a whole. This method will therefore define a single landmark for each of the clusters which could act as a beacon for navigation to the heart of the district. The second method places landmarks at cluster intersections wherever more than two districts intersect. This method results in fewer landmarks than simply using the centroids but those landmarks which are created appear in the areas of the space where a number of clusters meet. We anticipate that this kind of location is where a stable reference point might help in navigation and so is

The current LEADS implementation first identifies the gateway nodes and forms paths between nearest neighbour pairs in adjoining districts and then within districts all the gateway nodes are linked together with a spanning tree. Considering dynamic usage information, we propose that the positions of node, path and gateway features should evolve according to both the frequency of access to individual items and the sequence in which the items are examined. Applications

We have applied LEADS to three different visualisation systems which were all locally available at the start of this research. These are: the FDP-Grapher tool for visualising three dimensional network structures [Benford95]; the VR-VIBE system for interactive visualisation and searching of document databases and 3

questionnaires. While the size of the sample was small (six people in total) and we recognise that further studies are required to enhance the reliability of our provisional results, we nevertheless believe that they provide some indication of the potential of our approach - especially when the results are considered together in terms of their overall pattern.

the World Wide Web [Benford95b]; and the Q-PIT 3-D scatter-graph visualiser [Benford95]. Given space constraints, we will only focus on one these, FDPGrapher, here. FDP-Grapher is a 3-D tool for visualising arbitrary network structures. The underlying visualisation approach is based on the Force Directed Placement (FDP) technique where the nodes of a network are treated as masses, the links as springs and their mutual effects are simulated until the whole structure settles into a stable configuration from some random initial state [Fruchterman91]. The resulting visualisation shows the network drawn in 3-D space such that densely interlinked groups of nodes are positioned closely together. FDP-Grapher might have many applications. One of our current applications is to visualise regions of the World Wide Web. Users of this application are able to see an overview of up to a hundred or so linked Web nodes defined by an initial position (specified by a WWW URL) and a link adjacency distance. They can navigate the resulting visualisation with six degrees of freedom; select nodes in order to either obtain summary information or to launch the Mosaic browser in order to inspect their contents; and grab nodes and reposition them in order to stretch out the visualisation.

The main observation from statistical data is that the rate of task completion on the second and third attempts amongst users of the space with enhancement was higher than for those without. Users of the space with enhancements seemed to complete the task almost trivially on the third attempt while those in the raw space still had problems. This would seem to imply that greater learning of object positions did take place with the aid of legibility features. Another interesting point is that the mean time taken to find individual objects was consistently smaller on all attempts for the users of the space with legibility enhancements. These users seemed to gain some immediate advantage from the legibility features, without the chance for learning to take place. Observational and questionnaire data suggested that the subjects using the space with legibility enhancements seemed to remember the positions of the objects more effectively. They commented that the main aid to remembering the position of the objects was the colour (and to a lesser extent shape) of the districts. Although these were the main features used as memory cues, from observation it was clear that other features were also being used. Subjects using the space without legibility enhancements would often show signs of remembering certain attributes of the region of space an object was in but not its definite position.

Ingram, plates 1 and 2, show an example of LEADS being applied to FDP-Grapher when visualising a network of 239 nodes (the example network actually used for experimentation). Ingram, plate 1 shows the basic visualisation with no additional legibility features. Ingram, plate 2 shows the same visualisation with all of the legibility features present. In this case, districts are distinguished by colour and shape, edges are large green planes, paths are hi-lited in white, major nodes are made larger and landmarks appear as red spikes. It is important to bear in mind that this images provide a distant perspective view and that the user is able to fly right into the centre of the visualisation, in which case they are surrounded by the nodes and links. Thus, what might appear to be a somewhat cluttered image from this distance becomes more open from closer in or even inside. Please note that LEADS allows individual legibility features to be dynamically switched on and off and to be combined in different ways.

On the general question of whether the subject felt disoriented in the space, the surprising answer is that the group without legibility enhancements felt less disoriented. This may indicate that the extra information added to the space could possibly add to initial overload of information. However, on the related question of whether the users felt most disorientation within the group of objects or outside it (i.e. gaining an overview) the subjects from the group with enhancements seemed more comfortable with the latter, a view of the whole space. We might surmise then that the divisions presented by the clustering into districts are of most use on the global scale of the space and that more effective local cues for orientation are necessary.

Experimentation: finding objects in a data visualisation

The aim of initial experimentation was to test the way in which subjects learned the layout of an information space which they entered and used repeatedly. Subjects were asked to complete a search task to find 5 objects in a randomly generated graph of 239 nodes on 3 separate occasions over 2 days. Separate experiments were carried out with different subject groups in spaces with and without legibility enhancements. The experiment only considered the addition of all five of the legibility features together and not the effects of individual features or smaller combinations of features. Two types of result were gathered from the experiment: statistical, from the times taken to find each object; and anecdotal from observation of the subjects and through

The questionnaire also explored how subjects learned the positions of target objects. In only one of these questions did the subjects show significant agreement in the answers given. Here both groups indicated that they had gained a good impression of the overall structure of the space. However, the questions where the answers did differ are more interesting. The first of these was asking if the subjects knew where the search targets were when they attempted the trial for the third and final time. The subjects using the enhanced space were much more positive that they had learned the absolute locations of 4

the objects than those without enhancements. The other two questions were testing whether the subjects knew the positions of the targets relative to each other. The subjects with legibility features were again more positive that they knew the relative direction of their next target object and that they had learned routes between the objects.

Goal: virtual environments to promote social encounters and collaborative activity

Accordingly, our first goal is to deploy insights from architecture and urban planning to explore how virtual settlements can be built which, following Jacobs’ aphorism quoted above, generate contact between individuals and, through this, provide opportunities for collaborative activity. Following [Hillier84] we examine different configurations of urban form to see how real settlements achieve this. However, to make this background directly impact upon the development of virtual environments, we need some appropriate computational techniques by means of which virtual environments could be built which have the requisite configurational properties. Hence our second goal has been to refine ’city building algorithms’ and to examine cities built using them with respect to their navigability and how they generate contact through simulation experiments. In this paper, we only have space to give an overview of our research in relation to these goals. Further details can be found in [Bowers95].

In summary, our experiment provides initial evidence for the effectiveness of applying our algorithms, derived from Lynch’s notion of legibility, to automatically enhance 3-D information visualisations. In the following section, we describe a second piece of research, which has adapted ideas from a different area of urban planning, to consider support for generating effective social virtual spaces.

Theory: the social logic of space

However we think about it, if individuals move in space and social interaction takes place through meeting, space patterns that constrain movement may intervene in the construction of social behaviour. [Penn94, p.85]

THE SOCIAL LOGIC OF CYBERSPACE

"Towns are mechanisms for generating contact." (Jane Jacobs). Our second thread of research, although it also draws on work which has been highly influential in architecture and urban planning, takes a different perspective on issues to do with the intelligibility and navigability of environments be they real or virtual. Following the work of Bill Hillier and Julienne Hanson, whose most well known text is The Social Logic of Space [Hillier84], we are interested in the configurational properties of settlements themselves and whether the local and global spatial order of settlements can influence their navigability and intelligibility, and conversely, whether virtual settlements can be designed which capitalise on what is known about real ones.

Imagine an environment consisting of a number of subspaces (rooms, say) connected to a central space (an entrance hall, say) such that the only way one could move from one sub-space to another was through the central space, there being no direct connections from one sub-space to another. Clearly, such a spatial organisation, whether manifested in a real building or in a virtual environment where collision detection algorithms prevent navigation between the sub-spaces, will directly constrain both the accessibility of the different spaces from each other and in doing so it will produce differentiations in the environment with respect to the frequencies of social encounter. Provided individuals need to move from one location to another, the opportunities for unplanned social encounter will be more likely in the central, connecting space. This is a crude example of how a spatial arrangement can promote social encounters of certain sorts and, naturally, such phenomena are exploited in the design of buildings and exterior spaces and guide the siting of sources of information (e.g. public noticeboards) and other resources (e.g. telephones, drinks machines).

In particular, a research strategy of this sort should be able to inform the development of collaborative virtual environments (CVEs) where multiple users share, interact and work within a common virtual environment [Benford95, Bowers96]. Much work within the field of computer supported cooperative work (CSCW) has emphasised the importance of mutual awareness between individuals to sustaining cooperative work. The coordination of work activity is often achieved through people being able to monitor the conduct of others and displaying their own activity so that it can be ’tuned into’ by others [Heath92]. This suggests that CVEs should enable their users to be mutually aware of each others’ activity in ways which draw upon the methods people ordinarily use [Benford96]. However, to even become rudimentarily aware of others in a shared virtual environment requires some form of contact or social encounter to occur between individuals.

Hillier and Hanson [Hillier84] show, through an analysis of hamlets and small villages in the south of France, that even such simple, unplanned, ’organic’ settlements manifest an ’underlying order’ which subtly concentrates social encounters in some places and not others, which makes some parts of a settlement more readily accessible than others, and which overall produces a configuration which aids navigability. They call this configuration the beady ring and claim that it can be found in a variety of different geographical settings 5

irrespective of building materials and the basic shape of constituent buildings. A beady ring consists of a set of open spaces of varying sizes, all inter-connected to form a major ring and other sub-rings, with each individual building fronting onto open space so that the settlement has an inner clump of buildings with a set of outer clumps around it [Hillier84, p.57-58]. This structure also has the property of there being at least two routes from any one building to any other. While this allows the settlement a ’high degree of permeability’, the connections between the open spaces will mean that most routes between buildings fronting onto different open spaces will pass through the ’major ring’ or central open space.

information. In a maze, the structure of space is designed to break the part-whole relationship as far as possible. [Penn94, p.88] To summarise: These analyses reveal real settlement forms to be differentiated spaces. Locales within a differentiated space vary in terms of their integration with the overall structure and with respect to each other. Actual settlements (especially ’organically’ evolved, non-planned ones) are differentiated but without any necessary bounding or compartmentalization. This permits the use of space to evolve to take advantage of the differentiations that the configuration supports but without any necessary form of ’apartheid’ between social activities. Analysis of settlements in terms of the lines of sight and access that their open spaces afford reveals that they produce gradations of accessibility and awareness. It need not be the case for an inhabitant or stranger to be prohibited from a part of the city to nevertheless make it more or less hard to access. The city configuration itself tends to make some parts more highly integrated than others. These variations between different locales in a settlement and the distinctiveness of each promote navigability.

As settlement forms grow, though, Hillier and Hanson argue that further organising principles become manifest. One of these is the deformed grid [Hillier 84, p.90. Penn94, p.90] where the open space structure is marked by more or less linear streets, often punctuated by large convex ’squares’. Penn and Dalton remark that such squares have a crucial role in the overall organization of the small town form: Urban squares are often thought of as relatively enclosed, but if we look at these in detail we find that they are characterised more by the length of lines of sight that pass into and through them. In fact, the lines of sight from each square [will tend to] intersect one another and provide a more or less continuous route structure in which it is never more than one or two changes of direction from one square to the next.

In principle, these subtle inter-relations between access, lines of sight, navigability and probabilities of social encounter can be exploited in the implementation of suitably designed or evolved virtual villages, towns and cities which could serve as CVEs. In this way, city (etc) metaphors for virtual environments may produce gradients of accessibility for information and computational resources distributed about them, rather than insist that accessibility is an ’all-or-none’ matter determined by the possession of a password or some other access key. Our argument is that this features may fall out naturally in the design or evolution of virtual settlement-like spaces as a product of their configurational properties.

In short, the existence of squares provides a ’central core’ to the town enabling sight, movement and access from one part to another. These configurational properties have immediate implications for the navigability of urban forms. Wayfinding from one locale to another in a town of this sort can very often be effectively organised around (primarily) movement from one square to another and then (secondarily), on leaving a square near the destination, from street to street until the destination is found. If squares become points of passage along routes in this way, this reinforces their status as arenas for encounters, as stopping off points, as places to seek help for wayfinding and (in later developments) as places to lodge ’attractions’ and so forth.

Implementation: Virtual City Builder

Hillier and Hanson [Hillier84, p.59-61] outline a computer program for simulating the spatial aggregation of buildings to make up settlements which manifest many of the properties of simple real settlements that they have identified. VCB (for Virtual City Builder) considerably extends this work so that a greater variety of settlement forms can be generated and output as scene definition files for 3D visualisation. VCB proceeds by repeatedly aggregating elements onto a start or ’seed’ element. Each element consists of a closed cell and an open cell joined together. So far VCB has aggregated elements onto a 2D surface with geometric squares as cells and elements being composed by a full facewise join of the two (an ’entrance’ can be regarded as lying along the joining face of the closed cell). VCB allows elements to aggregate randomly, requiring that each new element added to the surface joins its open cell full facewise onto at least one other open cell. The location of the closed cell of the element is randomly selected from those sites available which are adjacent to

Observations like this enable Hillier and Hanson and those who follow them to develop an alternative to Lynch’s views on what makes for an ’intelligible’ form. We believe that one of the simplest ways of defining the term [intelligibility] is in the relationship between local and global properties of space. A system is intelligible to the degree that what you see immediately around you gives a good guide to where you are in the whole system. This is both a dynamic and a relational concept - as you move through a city, what you see is constantly changing, even though the whole city remains almost completely unchanged. In an intelligible area your global location is essentially predictable from local 6

the element’s open cell. However, the random selection of the closed cell location can be weighted by how many facewise neighbouring closed cells each available site has. Changing the weights from an even distribution will make VCB more or less likely to aggregate closed cells when they adjoin other closed cells. If closed cell selection is highly weighted towards those sites which already have, say, one closed cell neighbour, virtual cities are generated which have many long straight streets with narrow terraces of closed cells and rarely an isolated closed cell with no neighbours. An even probability distribution tends to produce forms which are more ’tangled’ with many short winding streets and small clumps of closed cells. Further constraints can be introduced on how VCB selects an element’s closed cell location from sites adjacent to the element’s open cell. One possibility is to exclude sites which have only a vertex-to-vertex join to one or more other closed cell(s). Without this condition, VCB’s virtual cities tend to have a structure whose open space structure contains no rings or ’circuits’. In graph theory terms, the open cells connect up to form a tree. On the other hand, if vertex-only joins between closed cells are disallowed, beady ring formations with varying numbers of rings or circuits can be generated. Indeed, as suggested in [Hillier84], varying the aggregation weightings determines the number of rings formed. The more the weighting distribution is skewed by increasing the relative probability that the aggregation of a closed cell will take place on a site with no closed cell neighbours, the more rings the formation has.

Figure 1: Virtual cities generated by VCB and visualised in DIVE 3.0. Evaluation: wayfinding and encounters among software agents

How might the approach to the generation of virtual, settlement-like environments outlined here be evaluated? There are a number of possibilities. First, as such environments are intended to be CVEs for multiple human users engaged in cooperative work, studies of some cooperative task in a VCB virtual city suggest themselves. However, just what that task should be requires a degree of specific application development to be undertaken first so that human performance could be realistically assessed. In the case of our work on LEADS, as it is clear that LEADS is devoted to facilitating information access, it is relatively straightforward to design appropriate experimental tasks. In the case of the VCB cities, their 'evaluation' by such means is much more problematic. The goal of providing an environment to support wayfinding and social action is, at this stage, too general to permit motivated evaluation studies with human users.

As we have mentioned, aggregation in VCB takes place on the 2D ’ground’ plane. Buildings can be generated for 3D visualisation by extruding the closed cell squares into boxes of varying (user-defined) heights and colours. Figure 1 shows a virtual city of 300 buildings with heights determined by their order of aggregation. This makes the ’seed cell’ into the tallest building with other heights ’decaying’ around it and so enables city visualisations with tall buildings tending to be (but not always) in central districts. This visualisation was made using the DIVE system (version 3.0) developed at the Swedish Institute of Computer Science [Fahlén93].

For these reasons we have preferred to conduct simulation experiments using software agents who embody simple wayfinding strategies. These experiments enable us to determine, for these agents, the properties of VCB-generated environments with respect to navigability and distribution of 'social' encounter. The simplicity of the agents is congruent with the simplicity of VCB's virtual cities (which are, after all, only boxes of open and closed space). In this respect, we believe that simulation experiments make for a fairer evaluation than evaluation with more complex agents (like humans). Penn and Dalton [Penn94] note that differences in the performance of simple agents across different cities are more likely to be due to configurational differences in the virtual cities than idiosyncrasies in agent design. We agree and wish to extend this argument to justify preferring software simulations over human experiments or 'field trials' at this stage of our research. Once our cities are more 7

Although we have designed a number of software agents, we shall only consider two in what follows (for more details, see [Bowers95]). Random walker sets off in a (unweighted) random direction from the start cell (unless the goal can already be seen) and makes random choices at junctions between neighbouring cells which have not been visited before. Longsight walker sets off from the start cell in the direction of the longest line of sight (unless the goal can already be seen) and chooses next steps at junctions in terms of which unvisited direction gives the longest line of sight. If two or more directions are equally long, a random choice is made. In other words, the direction down which ’longsight’ can see the furthest is selected from amongst those not already visited. A single experiment comprises the two walkers undertaking a 1000 trials (with a randomly selected start and goal cell each trial) in a number of different virtual cities. The mean performance for each walker in each city is then taken to comprise a measure of the navigability of that city for that walker.

detailed and have a closer tie to applications, then we feel we will be able to move to human evaluation. The simulation experiments we have done involve placing a wayfinding agent at a randomly selected open cell as its starting location with the task of finding another randomly selected open cell, the goal. Agents are given a ’memory’ of the open cells they have visited so far. Moving to an adjacent open cell takes one time unit or ’step’, as does backtracking to the last ’choice point’ if no unvisited open cells are available. The agents are assumed to be able to ’see’ in straight lines in the direction they are going (north, south, east or west) along the open space structure of the city. As soon as they can see the goal cell (i.e. a straight north-south or east-west connects them with the goal state), they can proceed straight towards it. Agents cannot see or walk through a closed cell. The total number of steps is recorded for an agent to reach the goal.

Figure 2: (a) A tree-like VCB city. (b) A labyrinth-like VCB city. (c) A beady ring VCB city. The dark squares show the closed cells with a ’notch’ indicating the entrance (and hence the joining face with the adjacent closed cell). See text for an explanation of the shading in Figure 2 (c).

In one experiment, two 100 element cities were compared (see Figure 2, a and b), one the result of a heavily weighted aggregation process so that closed cells would tend to adjoin just one neighbour (a tree-like city), one which did not weight the aggregation process (a more labyrinth-like city). Analysis of the data revealed a clear interaction between city form and walker strategy. The longsight walker reached the goal in 27% fewer steps than the random walker in the ’treecity’, while the random walker reached the goal in 9% fewer steps in the ’labyrinth-city’. ’Longsight’ finds the tree-city more navigable larger because its strategy of following long lines of sight is well suited to a virtual city which contains a number of long lines of sight, while ’random’ is likely to waste time meandering along and across large open spaces. In contrast, longsight can overlook goals which are hidden or positioned at the end of dead-ends in the labyrinth-city, while random is exhaustive in searching through a region within the city. This can give random an advantage for navigating cities with many short, winding streets and dead-ends.

Further experiments have been conducted on beady ring formations and studying the distribution of encounters between agents in different cities. Encounters between agents are investigated by studying the distribution of probabilities that two or more agents might find themselves on the same open cell in the same time unit. If we assume that agents make their journeys independently, calculating the number of times each open cell is visited in the 1000-trial experiments we have discussed gives an indication of where the cells are which will be ’dense’ with such chance encounters. Our analyses [Bowers95] reveal complex interactions between agent strategy and city type with particular city features ’attracting’ many encounters. Figure 2 (c) shows the distribution of encounter density across a beady ring city with the most dense open cells being shaded in the darker greys. This has been plotted for a city inhabited by longsight walkers. Note that the locations which give rise to the higher probabilities of encounter lie at the mouths of dead-ends, at junctions and along long lines 8

of sight. The corners of circuits (rings) are also prominent if they are junctions too.

CONCLUSION

This paper has explored how previous research from the discipline of urban planning into the relationship between real world urban environments and the human processes of navigation and social encounter can be adapted to the design of virtual environments. The discussion has been presented in two parts, each drawing on a different tradition from urban planning. Each part has proposed algorithms for automatically generating or enhancing a variety of shared virtual spaces and has presented the results of experimental work to support these proposals.

In summary, the simulation experiments show a number of interactions between agent strategy, city form and encounter probability. This indicates that the configurational properties of a city, as well as the wayfinding strategy used, can influence navigability and that the two factors interact. Just as there are ’horses for courses’ so there are ’wayfinding strategies for virtual cities’. These results give us some reason to believe that VCB does indeed generate virtual cities with systematically varying and controllable properties. By manipulating the aggregation algorithms one can derive a variety of forms which not merely look different subjectively on inspection of a plan view (compare the different parts of Figure 2) but actually lead to differences in navigability. If there were no differences between different forms in terms of their navigability or no interactions with different agent-types, we would not have evidence for the potential utility of VCB for generating city-like forms with properties akin to those of real cities. Furthermore, if the probabilities of encounters were evenly distributed across the open cells of a VCB form rather than systematically linked to different features (like junctions and long ’streets’), we would not be able to use VCB for generating cities which contain within them differentiations with respect to accessibility and encounter-probability.

The first major outcome of this work has been to emphasise the potentially rich cross-fertilisation that might take place between the disciplines of urban planning and the design of virtual environments. Thus, in addition to sociology and psychology, urban planning and architecture can now be seen as making a major future contribution to the endeavour of constructing this technology. Of course, recognition of the potential of this particular inter-disciplinary relationship is not particularly new. Architectural walk-throughs and planning scenarios have previously been proposed as potential applications of virtual reality, with a particular focus on how virtual reality technology (typically single-user) might support planners and architects in designing for the real world. In contrast, the focus of our discussion has been the reverse; how can urban planning and architecture contribute to the design of shared virtual spaces in their own right (i.e. virtual spaces with no physical counterpart)? In addition, some notable virtual reality researchers have backgrounds in architecture and planning [Benedikt91, Novak91]. However, previous research has tended to recognise the general relevance of the one field to the other or to use one to inspire ’manifestos’ for the other. We would argue that our work is possibly the first to make a sufficiently direct link between the two disciplines as to derive specific computational algorithms directly from planning principles and theories of urban form.

However, now that is has been shown that VCB can generate a variety of city forms with some desired properties, one can begin to develop applications which utilise such cities as CVEs in a principled way. For example, [Penn94] presents data suggesting that the length of a line of sight does indeed influence human walkers as they undertake journeys. For applications where humans might be undertaking a goal-directed search of a virtual city, then we could parameterise VCB’s algorithms to generate cities like those in part (a) of Figure 2 as those seem to be the most navigable for such agents embodying such strategies. Similarly, the results of the simulation studies allow us to advise on how access to information and computational resources might be distributed in a city-style CVE. For example, one might want to site access points to tools for the support of meetings close to locations where encounters are quite probable. In a city-style CVE, a ’virtual meeting room’ need never be far away from places where chance encounters are most likely to occur. Finally, [Bowers95] argues that, for many purposes, a beady ring formation like Figure 2 (c) is the most effective compromise between a city equally well navigable by a mix of wayfinding strategies while generating many systematically distributed ’social contact points’. Indeed, Bowers speculates that this might account for the ubiquity of the beady ring in many existing settlement forms.

The second outcome of this work has been a set of algorithms, backed up with experimental evaluation, to assist in the construction of virtual environments. These algorithms address both the enhancement of existing 3D visualisations through the automatic introduction of legibility features (i.e. the creation, emphasis and placement of virtual districts, landmarks, edges, paths and nodes) and also the creation of virtual cities with different underlying structures affecting navigation and opportunities for social encounter. Early evidence suggests how both approaches may be of benefit. Perhaps the most urgent question to be asked about the work described in this paper is "what is the relationship between the two approaches that have been followed?". In seeking to answer this question, we will briefly consider both the similarities and differences between the work described in parts 2 and 3. The two approaches are similar in terms of their general goals. They both 9

draw inspiration from urban planning, they both address the problem of navigating potentially large virtual spaces, they have both proposed algorithmic techniques for generating such spaces and they have both conducted experimental work to assess these algorithms. However, there are also notable differences between them. The first difference is that Lynch’s work adopts a cognitive orientation, focused on how an individual constructs a personal mental map of a virtual space, whereas Hillier and Hanson’s work emphasises the configurational properties of the environment itself and of the social relations environments afford. While these differences are substantive and lead to rich disputes within architecture and urban planning, it does not seem to us that either approach need exclude the other as a source of inspiration for the development of virtual environments. The integration of the algorithms we have derived may well be possible with various legibility features being introduced through the LEADS algorithms into a space produced by VCB. Indeed, [Bowers95] explores the possibility of using VCB in information visualisation by allowing ’semantic rules’ to influence the spatial aggregation process. In such cases, the legibility layer provided by LEADS could work upon an existing visualisation provided by VCB. That, the two approaches can be so readily combined should also reassure developers that it is not in strictness necessary for VR researchers drawing on theories of architecture and urban form to inherit the disputes within those disciplines! The second difference concerns the nature of the experimentation that has been conducted to date. Evaluation of LEADS has focused on trials with a small number of human subjects whereas evaluation of VCB has been based on larger numbers of simulation experiments where software agents embodying simple wayfinding strategies. As argued above, this reflects the relative proximities of the two approaches to specific applications. LEADS is directly concerned to provide support for specific information visualisation applications. This makes it reasonable to design exact experimental tasks and conduct studies with human subjects. VCB is more abstract, exploratory and concerned with general principles of CVE design. In our view, simulation studies are more reasonable at this level. This comparison suggests to us that different evaluation methods have varying utility at different stages of the unfolding of system development. Having summarised the similarities and differences between the two approaches, the most immediate direction for future work would seem to be in integrating the two. Current effort is being devoted to supporting the semantic aggregation of cells when generating a space within VCB and then using LEADS algorithms to provide additional legibility features within this space. We are considering various specific application areas where such augmented visualisations may be of use. Our intention then is to conduct studies

involving real human participants in VCB generated spaces. Current research CVEs are capable of supporting in the region of ten simultaneous users. However, one might expect to see research platforms capable of supporting a hundred or so users within the next year, making such experimentation a more interesting and likely prospect. Future work might also involve a closer consideration of the relationship between the two fields of urban planning and collaborative virtual environments. In particular, we propose that virtual environments might provide a kind of laboratory for examining theories within architecture and urban planning as well as a means of exploring the consequences of particular design choices. If the benefit of cross-fertilising VR research with studies of the built environment is experienced by both fields, then we can look forward to an exciting period of research and technology development. REFERENCES

[Benedikt91] Benedikt, M. (1991). A cyberspace manifesto. In M. Benedikt (ed.) Cyberspace: The first steps. Cambridge, MA: MIT Press. [Benford95b] Benford, S., Snowdon, D., Greenhalgh, C., Ingram, R., and Knox, I., (1995) VR-VIBE: A Virtual Environment for Co-operative Information Retrieval, Computer Graphics Forum, 14, 3, 1995, NCC Blackwell. [Benford95] Steve Benford, John Bowers, Lennart Fahlén, Chris Greenhalgh, John Mariani and Tom Rodden, Networked Virtual Reality and Cooperative Work, Presence, MIT Press (in press). [Bowers95] John Bowers (1995) The social logic of cyberspace or The interactional affordances of virtual brutalism. In A. Bullock and J. Mariani (eds.) COMIC project deliverable 4.3, available from Lancaster University, UK. [Bowers96] Bowers, J., O'Brien, J. and Pycock, J. Talk and embodiment in collaborative virtual environments. To appear in the Proceedings of CHI'96. New York: ACM. [Fahlén93] Fahlén, L. E., Brown C. G., Stahl, O., Carlsson, C., A Space Based Model for User Interaction in Shared Synthetic Environments, in Proc.InterCHI’93, Amsterdam, 1993, ACM Press. [Fairchild93] Fairchild, K. M., Information Management Using Virtual Reality Based Visualisations, in Virtual Reality: Applications and Explorations, Wexelblat, A., (ed), Academic Press, Cambridge, MA, 1993, pp. 45-74. [Fruchterman91] T. M. J. Fruchterman, T. M. J. and Reingold, E. M., Graph Drawing by Force Directed Placement, Software Practice and Experience, Vol 21 (11), 1129-1164, November 1991. 10

[Heath92] Heath, C. & Luff, P. (1992). Collaboration and control: Crisis management and multimedia technology in London Underground Line Control Rooms. Computer Supported Cooperative Work, 1 (1/2), 69-94. [Hillier84] Hillier, B. & Hanson, J. (1984). The social logic of space. Cambridge: CUP. [Ingram96] Rob Ingram and Steve Benford, The Application of Legibility Techniques to Three Dimensional Information Visualisations, available from the authors (address at front of paper). [Lynch60] Lynch, K. (1960). The image of the city. Cambridge, MA: MIT Press [Nielsen90] Nielsen, J., The Art of Navigating Hypertext, CACM, 33, 3, March 1990, pp. 297-310, ACM Press. [Novak91] Novak, M. (1991). Liquid architecture. In M. Benedikt (ed.) Cyberspace: First steps. MIT Press. [Passini92] Passini, R., Wayfinding in Architecture, Van Nostrand Reinhold, 1992. [Penn94] Penn, A. & Dalton, N. (1994). The architecture of society: Stochastic simulation of urban movement. In G. N. Gilbert & J. Doran (eds.) Simulating societies: The computer simulation of social phenomena. London: UCL. [Zahn] Zahn, C.T., Graph-theoretical Methods for Detecting and Describing Gestalt Clusters, IEEE Transactions on Computers, C 20, 68-86.

11

Ingram, plates 1 and 2: LEADS applied to FDP- Grapher: (a) without legibility and (b) with legibility

12

Suggest Documents